Learning for Visual Data Compression


Dong Xu
University of Sydney

Guo Lu
Beijing Institute of Technology

Shan Liu

Shenlong Wang

Raquel Urtasun
Uber ATG / University of Toronto

Radu Timofte
ETH Zurich

Guest Speakers

Ren Yang
ETH Zurich



Schedule (EDT)

10:00 - 10:05 Introduction to the speakers and our tutorial Dong Xu
10:05 - 10:55 Deep image compression [slides] Ren Yang
10:55 - 11:45 Deep video compression [slides] Guo Lu
11:45 - 12:15 Break --
12:15 - 13:00 Deep point cloud compression [slides] Shenlong Wang
13:00 - 13:30 Visual data compression standards [slides] Shan Liu


In this tutorial, we will introduce the recent progress in deep learning based visual data compression, including image compression, video compression and point cloud compression. In the past few years, deep learning techniques have been successfully applied to various computer vision and image processing applications. However, for the data compression task, the traditional approaches (i.e., block based motion estimation and motion compensation, etc.) are still widely employed in the mainstream codecs. Considering the powerful representation capability of neural networks, it is feasible to improve the data compression performance by employing the advanced deep learning technologies. To this end, the deep leaning based compression approaches have recently received increasing attention from both academia and industry in the field of computer vision and signal processing.

Specifically, we will first introduce the end-to-end learning based image and video compression methods and discuss the current benchmark results. Then, we will provide detailed introductions for the latest standard procedures for learning based image or video compression approaches, such as JPEG AI, JVET NNVC and IEEE FVC. After that, we will discuss the recent work on learning based point cloud compression and analyze several widely used point cloud processing methods. Finally, we will discuss the limitations of the current learning based compression methods and the future research directions, like video compression for machines. In summary, our tutorial will cover both latest works from the academic community and the standardization progress in industry, which will help the audiences with different backgrounds better understand the recent progresses in this emerging research area.

Tutorial Outline

  1. Standard Activities of learning based Image and Video Compression
    a) Brief introduction to standards involving learning based image and video compression.
    b) Latest progress on learning based image and video coding tools in various standards.
    c) Discussion and Benchmark Results

  2. End-to-end Learning based Image and Video Compression
    a) Brief introduction of classical image and video compression frameworks
    b) Learning based image compression
    c) Learning based video compression
    d) Rate distortion optimization techniques for learned image and video compression
    e) Benchmark results and Discussions

  3. Learning based Point Cloud Compression
    a) Existing work for traditional point cloud compression
    b) Learning based static point cloud geometry compression
    c) Learning based dynamic point cloud geometry compression
    d) Learning based point cloud attribute compression

  4. Discussion and Future Directions
    a) Limitations of the current learning based approaches
    b) Visual data compression for machines
    c) Visual data compression for robotics and self-driving
    d) Open source projects